Intrinsic Geometry of Hard Disk Clusters

Fabi\'an Henry Vilaxa; Jos\'e Ayala Hoffmann

REVIEW 3 major objections 5 minor 19 references

Intrinsic Geometry of Hard Disk Clusters

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The minimum perimeter of a cluster of five unit disks is 10+2π, attained exactly by contact classes X10, X11, and X12.

desk verdict Likely-true solution of the five-disk perimeter problem, but the written proof is conditional on an unpublished computational certificate and an unproved connected-graph reduction. read the letter →

arxiv 2608.06513 v1 pith:YLENHPQ2 submitted 2026-08-06 math.MG

José Ayala Hoffmann , Fabián Henry Vilaxa This is my paper · ORCID

classification math.MG MSC 52C1552A4049Q1051M25

keywords diskpackingharddisksperimeterminimisationconvexhullcontactgraphsrollingspaceintrinsicHessianpenny

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The minimum perimeter of a cluster of five unit disks is $10+2\pi$. No arrangement does better, and the proof shows that the bound is attained exactly by three realised contact classes. The paper builds a general calculus for hard disk clusters: configurations are grouped into fixed hull-and-contact classes, where the perimeter has a smooth local formula and the allowed infinitesimal motions form the rolling space. Within each class it derives first-order criticality conditions, a first-order descent test based on hull leaves, and a second-order intrinsic Hessian that distinguishes instability, flat degeneracy, and rigidity. Applying this calculus to five disks, it enumerates thirteen possible contact graphs, excludes ten of them, and identifies the three surviving classes as the complete set of minimizers. The result matters because the minimum perimeter problem was previously solved only up to four disks.

What carries the argument

The load-bearing object is the fixed hull and contact class $X=F\cap C(G)$, where $F$ fixes the set and cyclic order of hull vertices and $C(G)$ fixes the realised contact graph. At a configuration $c$, each active contact $\{i,j\}$ contributes the normal vector $u_{ij}=(c_j-c_i)/\|c_j-c_i\|$, and the rolling space $\operatorname{Roll}(c)=\ker A(c)$ consists of infinitesimal motions that preserve every contact to first order, with $A(c)$ the contact operator. Class criticality is the multiplier equation $\nabla\widetilde{\operatorname{Per}}(c)=A(c)^\top\$\lambda$$. The second variation is computed as the intrinsic Hessian $H_{\mathrm{intr}}=Z^\top(H_{\mathrm{hull}}+H_{\mathrm{cont}})Z$, where $Z$ spans $\operatorname{Roll}(c)\cap R(c)^\perp$, $H_{\mathrm{hull}}$ is assembled from transverse blocks $M_{pq}=\|c_q-c_p\|^{-1}(I-t_{pq}t_{pq}^\top)$ for hull edges, and $H_{\mathrm{cont}}$ from contact curvature blocks $K_{ij}=-\lambda_{ij}(I-u_{ij}u_{ij}^\top)/2$. The sign of $H_{\mathrm{intr}}$ on the reduced rolling space decides whether a class-critical configuration is second-order unstable, flat degenerate, or rigid modulo rigid motions.

What would settle it

Take the exact symbolic coordinates recorded in Appendix B for the parallel-leaf realisations of classes $X_2, X_3, X_5, X_6, X_8, X_9$, form the contact operator $A(c)$ and the reduced rolling space $\operatorname{Roll}(c)\cap R(c)^\perp$, and compute the intrinsic Hessian $H_{\mathrm{intr}}=Z^\top H Z$ in interval arithmetic. The paper certifies a negative eigenvalue for each; a single class for which the certified interval for the smallest eigenvalue contains zero or positive values would refute the exclusions and reopen the five-disk problem.

Watch

Extended reading notes

Core claim

The paper's central result is Theorem 6.1: among all configurations of five non-overlapping unit disks, the minimum of $\operatorname{Per}(c)$, the perimeter of the convex hull of the cluster, equals $10+2\pi$. The minimum is attained precisely by configurations in the realised contact classes $X_{10}$, $X_{11}$, and $X_{12}$, and by no others. In $X_{10}$ and $X_{11}$ the minimisers are not isolated: they admit local perimeter-preserving admissible flexes of dimension one and two respectively, while the minimiser in $X_{12}$ is rigid modulo rigid motions, meaning its rolling space consists only of infinitesimal rigid motions. On the way, the paper develops a variational calculus in which the realised contact normals and the cyclic hull order determine, respectively, the admissible first-order cone and the local form of the perimeter functional.

Load-bearing premise

The theorem depends on the unpublished computational certificate [AHV26], which identifies all parallel-leaf realisations in classes X2, X3, X5, X6, X8, X9 and certifies a negative intrinsic Hessian eigenvalue for each; it also assumes, without proof, that any disconnected minimizer can be replaced by an incident connected configuration with no larger centre-hull perimeter.

Editorial extensions

If this is right

Any five-disk minimizer has centre-hull perimeter exactly $10$, since $\operatorname{Per}(c)=\operatorname{Per}(P(c))+2\pi$.
The three minimising classes are fully classified: $X_{10}$ and $X_{11}$ contain local one- and two-parameter perimeter-preserving families, while $X_{12}$ is an isolated minimiser modulo rigid motions.
The first-order hull-leaf test and the second-order intrinsic-Hessian test give a general exclusion scheme that does not require solving a global optimization problem; the same scheme can be applied to any fixed hull and contact class for larger $n$.
The five-disk problem reduces to checking thirteen connected penny graphs, of which ten are excluded; this confirms the enumeration count $a(5)=13$ of connected penny graphs on five vertices.
Since every hull edge in the surviving classes $X_{10}$ and $X_{11}$ has length two, the perimeter is constant on those classes, so the minimum is achieved by whole flexing families rather than by a single shape.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same machine should extend to six disks: the connected penny-graph count grows to 46, and the paper's reductions—fixed hull classes, hull-leaf exclusion, and intrinsic Hessian diagnostics—are all finite, so the main new cost is computational rather than structural.
The flat-degeneracy criterion suggests a tighter link to rigidity theory of sticky disks: a class whose hull edges are all contacts is locally perimeter-constant, so perimeter minimality and rigidity become complementary rather than competing properties.
A direct testable extension is to run the interval-arithmetic certificate on the six-disk enumeration and look for a surviving class with positive semidefinite intrinsic Hessian but no hull leaf; such a class would be the first candidate for the next minimum that is neither trivially flat nor rigid.
The method's reproducibility hinges on the unpublished certificate [AHV26]; posting it would turn the five-disk theorem into a fully checkable proof.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The headline: this paper gives the first serious attack on the five-disk minimum perimeter problem and states the result as 10+2π, attained in exactly three contact classes. I think the result is probably correct, but as submitted it is not fully auditable.

What is genuinely new: the variational framework built around the rolling space, hull leaf obstruction, and intrinsic Hessian is a clean and useful assembly of rigidity and optimization ideas. The hull leaf theorem (4.2) is a nice first-order exclusion tool, and the analytic exclusions for X1, X7, and X4 are solid. The classification of the surviving classes, with flat perimeter-preserving flexes in X10 and X11 and rigidity of X12, is exactly what the theory should produce. The paper is also honest about its computational dependence: Appendix B describes the interval-arithmetic certificate rather than hiding it.

The soft spots are precisely where the stress-test note puts them. The exclusions of X2, X3, X5, X6, X8, and X9 all rest on [AHV26], which is "available upon request" and not provided. A referee cannot check the completeness of the parallel-leaf enumeration or the negative-pivot certificates. The X4 boundary also passes into X8 and X9, so it inherits the same dependence. That is not necessarily fatal, but it is load-bearing.

Second, Section 6.1 asserts without proof that any disconnected minimizer can be replaced by an incident connected configuration with no larger center-hull perimeter. This is probably true and likely easy to prove, but as written it is an unproved reduction.

The X13 exclusion is terse but acceptable: the paper excludes the degenerate horizontal case and non-convex hulls via the hull leaf theorem.

Overall, the mathematics is coherent and the gaps are addressable. I would send this to a serious referee rather than desk-reject. I would not cite the five-disk theorem in my own work until the certificate is public and the connected-graph reduction is proven, but I would bring it to a reading group focused on packing and rigidity.

Referee Report

3 major / 5 minor

Summary. The paper develops a variational calculus for the perimeter of the convex hull of hard disk clusters, based on fixed hull-contact classes, the rolling space of contact-preserving infinitesimal motions, a first-order hull-leaf obstruction, and an intrinsic Hessian on the reduced rolling space. As an application it proves (Theorem 6.1) that the minimum perimeter of a configuration of five unit disks is 10+2π, attained exactly in contact classes X10, X11, and X12, with local perimeter-preserving flexes of dimensions one and two in the first two classes and rigidity modulo rigid motions in the third. The proof combines analytic exclusions (X1 and X7 by a strict-concavity argument; hull-leaf classes by Theorem 4.2; X4 by an explicit one-parameter descent family) with computational certificates for the parallel-leaf realisations in classes X2, X3, X5, X6, X8, and X9, supplied by the referenced but unpublished file [AHV26].

Significance. If fully validated, this is a meaningful advance: it is the first exact determination of the minimum hull perimeter for n=5, and the proposed intrinsic framework (rolling space, first-order leaf obstruction, intrinsic Hessian diagnostics) may be reusable for larger n. The analytic parts I checked are sound: the X1/X7 concavity exclusion, the hull-leaf obstruction, and the X4 descent family are correct and clearly presented. The four-disk spectral prototypes in Appendix A usefully illustrate the distinction between rigidity, flat degeneracy, and second-order instability. The reported value 10+2π is consistent with the numerical literature. However, the proof of the main theorem is not self-contained as published, because a load-bearing reduction and the decisive computational certificate are not included.

major comments (3)

[§6.1] The assertion that 'any disconnected minimising configuration can be replaced by an incident configuration with an additional contact and no larger centre hull perimeter' is load-bearing, because it restricts the search to the 13 connected penny graphs, but no proof or reference is given. Adding a contact can change the hull and may increase the perimeter; a rigorous argument (or a precise citation to a known lemma) is required. Without this, the finite enumeration is incomplete.
[§6.2 and Appendix B] The exclusions of the hull-leaf classes X2, X3, X5, X6, X8, and X9 rest entirely on the unpublished certificate [AHV26], which is listed as 'available upon request' rather than provided. The certificate is asserted to (i) identify all parallel-leaf realisations in each class, and (ii) provide a certified negative intrinsic Hessian eigenvalue via interval arithmetic. The X4 exclusion ends at boundary configurations belonging to X8 and X9, so it inherits the same dependency. A referee cannot audit the completeness of the enumeration, the construction of the reduced rolling-space basis Z, or the interval arithmetic bounds described in Appendix B.2. The Python script, the dataset, and the verification logs must be made publicly available, and the pipeline described in enough detail to be independently reproduced.
[§6.2, proof of Theorem 6.1] The sentence 'For X10 and X11, every hull edge is an active contact of length 2, so Per(P(c)) = 10' is incomplete as written. If the hull had only k<5 vertices, all of whose edges are contacts of length 2, the perimeter would be 2k, which is strictly less than 10 and would contradict the claimed lower bound. The proof should explicitly state that the unique hull class in X10 and X11 has all five centres as hull vertices, or cite the certificate for that fact. This is a small but real gap in the written proof of the central theorem.

minor comments (5)

[§1] There is a typo: 'depend stronlgy' should read 'depend strongly'.
[Appendix B.2] The description 'Integer-pivoted Gaussian elimination yields a basis Z for Roll(c)∩R(c)⊥' is terse; a reference to the specific algorithm or a brief derivation would improve reproducibility.
[Table 1] The eigenvalue interval for X5, [−12.204,−7.2100], is much wider than the other intervals; a footnote explaining whether this reflects multiple realisations or numerical issues would help the reader.
[§6.2, X4 paragraph] The notation c0(t) is introduced, but the exact set of five active contacts is not listed; a short explicit statement of which contacts are preserved would make the verification easier to follow.
[References] The reference [AHV26] should be updated to include a repository URL or arXiv identifier, rather than only 'available upon request', so that the computational claims can be checked.

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional or fitted circularity: the minimum 10+2π is not assumed as an input, and the proof's analytical core is independent. The only self-support issue is the unpublished self-cited verification file [AHV26], which makes the finite leaf-locus exclusions unauditable but does not make the derivation circular.

full rationale

The derivation of Theorem 6.1 does not define or fit the target value 10+2π. The proof reduces to 13 connected penny-graph classes by an external enumeration benchmark (OEIS A085632; plantri/nauty), excludes X1 and X7 by a proven concavity argument, excludes the leaf classes by the proven hull-leaf obstruction (Theorem 4.2) together with negative-intrinsic-Hessian certificates, handles X4 by an explicit admissible descent whose boundary terminates in already-excluded classes, and computes the survivor perimeters directly from active contact/hull-edge lengths (Per(P(c)) = 10 for X10, X11, X12). None of these steps assumes the conclusion, and no fitted parameter is renamed as a prediction. The main caveat is that the exhaustive leaf-locus enumeration and negative-Hessian certification for X2, X3, X5, X6, X8, X9 are delegated to the authors' own 'Available upon request' file [AHV26], so the finite exclusions cannot be fully audited from the published text; this is a verifiability and self-citation concern, not a circular reduction. The unproved connected-graph reduction in §6.1 is an independent gap in the proof, again not a circularity. Overall, the central claim has independent content and is not forced by definition, by a fitted parameter, or by a self-citation chain.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The paper introduces no fitted constants; its mathematical content is a variational calculus plus a finite case analysis. The main burden is the unshipped computational certificate and the unproved connected-graph reduction.

assumptions (6)

standard math Planar Steiner formula: Per(P⊕D)=Per(P)+2π
Used in Proposition 2.2 to reduce cluster perimeter to centre hull perimeter plus 2π.
standard math diam(K) ≤ Per(K) for planar compact convex sets
Used in Theorem 2.4 to bound minimizing sequences.
standard math Harborth's edge bound |E| ≤ floor(3n - sqrt(12n-3)) for penny graphs
Used in §6.1 to bound the number of contacts for n=5, giving |E| ≤ 7.
domain assumption The enumeration of connected penny graphs on five vertices by plantri/nauty is complete and yields 13 isomorphism classes (OEIS A085632)
Load-bearing for the finite reduction in §6.1; matches an external sequence but the pipeline is not shipped.
ad hoc to paper Any disconnected minimizer can be replaced by an incident configuration with an additional contact and no larger centre hull perimeter
Asserted in §6.1 without proof; justifies considering only connected contact graphs.
domain assumption The [AHV26] interval-arithmetic certificate is correct
Underlies the exclusions of X2, X3, X5, X6, X8, X9 and the survival checks; code and data are not available.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Intrinsic Geometry of Hard Disk Clusters." pith.science (2026). https://pith.science/paper/YLENHPQ2

@misc{pith2026260806513,
  author       = {Pith},
  title        = {Pith review of: Intrinsic Geometry of Hard Disk Clusters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YLENHPQ2}},
  note         = {Machine review of arXiv:2608.06513}
}

abstract

Put $n$ identical coins on a table with no two overlapping. Which arrangement makes the perimeter of the convex hull of the cluster as small as possible? Despite its elementary statement, the solution of this problem is known only up to four disks. We produce a calculus for hard disk clusters of arbitrary finite size, providing class criticality conditions, first order descent tests, and second order spectral criteria for perimeter minimisation. A central difficulty is that the perimeter formula changes with the hull combinatorics, while the admissible first order geometry changes with the realised contacts. Our approach is guided by the principle that the realised geometry intrinsically determines both the local form of the functional and the admissible motions. As an application, this article takes the first step beyond four disks by providing a solution for the five disk case. The minimum perimeter is $10+2\pi$, attained in exactly three realised classes. Two admit perimeter preserving flexes, of dimensions one and two, while the third is rigid modulo rigid motions. The first order theory provides pruning criteria, and the reduced admissible space together with its intrinsic Hessian distinguish rigidity, second order instability, and perimeter flat degeneracy.

Figures

Figures reproduced from arXiv: 2608.06513 by the authors.

**Figure 1.** Complete list of connected contact graph types for n = 3 and n = 4. For these cases, the minimum perimeter problem can be checked by hand using elementary planar geometry. The list sizes are 13 for n = 5, 46 for n = 6, 162 for n = 7, and 715 for n = 8. This growth suggests the computational character of the problem, once a finite reduction has been established. These values are entries of the connected penny graph s… view at source ↗

**Figure 2.** The cluster hull is the Minkowski sum h(c) = P(c)⊕D, so Per(c) = Per(P(c))+2π. The centre hull, realised contacts, and contact graph are shown. Note the exterior angles of the centre hull sum to 2π. Theorem 2.4 (Existence of minimisers). For each fixed n ≥ 1, the perimeter functional Per : Dn → R attains its minimum on Dn. Proof. Let {c m} ⊂ Dn be a minimizing sequence for Per. Using translation invariance, we may a… view at source ↗

**Figure 3.** Realisations of the same contact graph with different cyclic hull data: left, (15)(54)(41); right, (12)(25)(54)(41). Contact graph edges are unordered, while cyclic hull data are recorded as ordered hull edges, up to cyclic shift and reversal. 3.2. Prescribed lengths versus realised contacts. Under bar–joint rigidity one starts with an abstract graph and prescribes the edge length constraints. For an edge {i, j}, t… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Geometry at a realised contact, showing the normal direction uij , the tangential direction u ⊥ ij , and perturbations of the centres. The first order non overlap condition is ⟨uij , δcj − δci⟩ ≥ 0, with equality corresponding to contact preservation. Suppose that c ∈ …

**Figure 5.** Figure 5: Realisations c1, . . . , c13 belonging to the fixed hull and contact classes X1, . . . , X13 displayed in the finite verification. configurations in which every active edge has length two and every non edge distance is strictly greater than two. This yields exactly thi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

[1]

Asimow and B

L. Asimow and B. Roth, The rigidity of graphs, Transactions of the American Mathematical Society, Vol. 245, No. 2, pp. 279--289, 1978

work page 1978
[2]

Gordian Unlinks

J. Ayala Hoffmann and J. Hass, Gordian unlinks, Revista Iberoamericana de Matem\'aticas, accepted. arXiv:2502.08499

work page Pith review arXiv
[3]

Ayala Hoffmann and F

J. Ayala Hoffmann and F. Henry Vilaxa, Computational verification files for Intrinsic Geometry of Hard Disk Clusters. Available upon request

work page
[4]

Baryshnikov, P

Y. Baryshnikov, P. Bubenik, and M. Kahle, Min-type Morse theory for configuration spaces of hard spheres, International Mathematics Research Notices 2014, no. 9, 2577--2592

work page 2014
[5]

Brinkmann and B

G. Brinkmann and B. D. McKay, Fast generation of planar graphs, MATCH Commun. Math. Comput. Chem. 58 (2007), 323--357

work page 2007
[6]

Carlsson, J

G. Carlsson, J. Gorham, M. Kahle, and J. Mason, Computational topology for configuration spaces of hard disks, Physical Review E 85 (2012), 011303

work page 2012
[7]

Connelly, Rigidity and Energy, Inventiones Mathematicae, Vol

R. Connelly, Rigidity and Energy, Inventiones Mathematicae, Vol. 66, No. 1, pp. 11--33, 1982

work page 1982
[8]

Connelly, S.\,J

R. Connelly, S.\,J. Gortler, and L. Theran, Rigidity for sticky discs, Proceedings of the Royal Society A, Vol. 475, No. 2222, 2019, 20180773. doi:10.1098/rspa.2018.0773

work page arXiv 2019

Show all 19 references

[9]

Fejes T\'oth, Lagerungen in der Ebene, auf der Kugel und im Raum, Grundlehren der mathematischen Wissenschaften, Vol

L. Fejes T\'oth, Lagerungen in der Ebene, auf der Kugel und im Raum, Grundlehren der mathematischen Wissenschaften, Vol. 65, Springer-Verlag, Berlin, 1953

1953
[10]

a tze f \

H. Groemer, Existenzs \"a tze f \"u r Lagerungen im euklidischen Raum , Math. Z. 81 (1963), 260--278

1963
[11]

Harborth, L\"osung zu Problem 664A, Elemente der Mathematik, 29, 14--15, 1974

H. Harborth, L\"osung zu Problem 664A, Elemente der Mathematik, 29, 14--15, 1974

1974
[12]

Kallrath, J

J. Kallrath, J. Ryu, C. Song, M. Lee, and D.-S. Kim, Near optimal minimal convex hulls of disks, Journal of Global Optimization, Vol. 80, No. 3, pp. 551--594, 2021

2021
[13]

B. D. McKay and A. Piperno, Practical graph isomorphism, II, J. Symbolic Comput. 60 (2014), 94--112

2014
[14]

Oler, A finite packing problem, Canad

N. Oler, A finite packing problem, Canad. Math. Bull. 4 (1961), 153--155

1961
[15]

Sch\"urmann, On Extremal Finite Packings, Discrete & Computational Geometry, Vol

A. Sch\"urmann, On Extremal Finite Packings, Discrete & Computational Geometry, Vol. 28, No. 3, pp. 389--403, 2002

2002
[16]

N. J. A. Sloane, editor, The On-Line Encyclopedia of Integer Sequences, published electronically at https://oeis.org, Sequence A085632
[17]

A. B. Hopkins, F. H. Stillinger, and S. Torquato, Densest local sphere-packing diversity: General concepts and application to two dimensions, Physical Review E 81 (2010), 041305

2010
[18]

Wegner, \"U ber endliche Kreispackungen in der Ebene , Studia Sci

G. Wegner, \"U ber endliche Kreispackungen in der Ebene , Studia Sci. Math. Hungar. 21 (1986), 1--28

1986
[19]

Whiteley, Rigidity and scene analysis, in Handbook of Discrete and Computational Geometry, J.\,E

W. Whiteley, Rigidity and scene analysis, in Handbook of Discrete and Computational Geometry, J.\,E. Goodman and J. O'Rourke (eds.), CRC Press, 1997, pp. 893--916

1997

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Asimow and B

L. Asimow and B. Roth, The rigidity of graphs, Transactions of the American Mathematical Society, Vol. 245, No. 2, pp. 279--289, 1978

work page 1978

[2] [2]

Gordian Unlinks

J. Ayala Hoffmann and J. Hass, Gordian unlinks, Revista Iberoamericana de Matem\'aticas, accepted. arXiv:2502.08499

work page Pith review arXiv

[3] [3]

Ayala Hoffmann and F

J. Ayala Hoffmann and F. Henry Vilaxa, Computational verification files for Intrinsic Geometry of Hard Disk Clusters. Available upon request

work page

[4] [4]

Baryshnikov, P

Y. Baryshnikov, P. Bubenik, and M. Kahle, Min-type Morse theory for configuration spaces of hard spheres, International Mathematics Research Notices 2014, no. 9, 2577--2592

work page 2014

[5] [5]

Brinkmann and B

G. Brinkmann and B. D. McKay, Fast generation of planar graphs, MATCH Commun. Math. Comput. Chem. 58 (2007), 323--357

work page 2007

[6] [6]

Carlsson, J

G. Carlsson, J. Gorham, M. Kahle, and J. Mason, Computational topology for configuration spaces of hard disks, Physical Review E 85 (2012), 011303

work page 2012

[7] [7]

Connelly, Rigidity and Energy, Inventiones Mathematicae, Vol

R. Connelly, Rigidity and Energy, Inventiones Mathematicae, Vol. 66, No. 1, pp. 11--33, 1982

work page 1982

[8] [8]

Connelly, S.\,J

R. Connelly, S.\,J. Gortler, and L. Theran, Rigidity for sticky discs, Proceedings of the Royal Society A, Vol. 475, No. 2222, 2019, 20180773. doi:10.1098/rspa.2018.0773

work page arXiv 2019